Accurate and timely prediction of crop yields is of crucial significance for ensuring food security, optimizing resource allocation and stabilizing market prices. The traditional yield prediction methods, which often rely on single-variable data modeling and neglect the temporal relationships between data, tend to lack stability and accuracy. To address this issue, a stacked model based on attention mechanisms and Long Short-Term Memory networks (StaMaLSTM) was developed for corn yield prediction by using the corn yield data of prefecture-level cities in  six provinces including Heilongjiang, Liaoning and Shandong from 2010 to 2022, combined with corresponding climate, vegetation, soil, and spatial topography data. Results show that the StaMaLSTM model achieves optimal performance at T=2 in multi-time-step prediction experiments. With the integration of the attention mechanism, the model attains a normalized root mean square error (NRMSE) of 16.53%, a mean absolute error (MAE) of 8.72, a mean absolute percentage error (MAPE) of 7.27%, and a coefficient of determination (R2) of 0.995, outperforming traditional LSTM and CNN deep learning models as well as SVM and RF machine learning models. When any single data category is removed, the prediction errors of StaMaLSTM, LSTM and CNN models increase slightly but remain relatively stable. In contrast, the SVM and RF models exhibited significant fluctuations in prediction results when data was missing, particularly with the absence of climate data, followed by soil data, while the impacts of vegetation and topographic data varied by model. These findings offer new pathways for crop yield prediction and provide important guidance for the prediction of other crops.

提前且准确地预测作物产量对于保障粮食安全、优化资源分配和稳定市场价格具有关键性意义。传统的产量预测方法主要基于单一变量数据建模，并常忽略数据间的时序关系，导致预测稳定性与准确性不足。为此，使用2010—2022年黑龙江、辽宁和山东等6个省份的地级市玉米产量数据，结合相应的气候、植被、土壤及空间地形数据，构建基于注意力机制和长短时记忆网络的堆叠模型(StaMaLSTM)进行玉米产量预测。研究发现：在玉米产量预测的多时间步长对比试验中，StaMaLSTM模型在时间步长T=2时表现最优；引入注意力机制的StaMaLSTM模型的归一化均方根误差NRMSE为16.53%、平均绝对误差MAE为8.72、平均绝对百分比误差MAPE为7.27%、决定系数R2为0.995，性能超过传统LSTM和CNN深度学习模型，以及SVM和RF机器学习模型；在移除任何一类数据后，StaMaLSTM、LSTM和CNN模型的预测误差均上升，但较为稳定。相比之下，SVM和RF模型在数据缺失情况下预测结果波动较大，尤其是缺少气候数据时影响最为显著，其次是土壤数据，而植被和地形数据的影响则因模型而异。为作物产量预测提供新途径，并对其他作物的预测具有重要指导意义。